I think it’s a tough call in this case. The Tesla is trying not to block an intersection where a light is green but there’s no room to clear the intersection on the other side due to traffic ahead. In fact an oncoming car is able to turn left in front of the Tesla because it waited.

Probably the law is that you should NOT enter the intersection in such a state, but the human nature would be to make a more nuanced judgement of “how bad would it be” to continue thru and possibly get stuck sticking out into the intersection for a bit until traffic moves again.

I would think - how long might I be stuck? Would it actual impede other traffic? Also factors like, am I’m late for work? Am I frustrated because traffic has been horrible or am I enjoying a Sunday drive?

Ego (Tesla’s name for the driving software) doesn’t get impatient. It can maximally obey all traffic regulations at all times if they code it to do so, even if human drivers behind might be laying on the horn.

This little clip really shows how much nuance is involved in designing these systems, and how doing the technically right thing can either be wrong or at least ambiguously right.

16:40 to 17:12

The driver in this case could see that the cars beyond the blue car were starting to move, and thus predict that he could cross a little bit early (betting on the space being available by the end of the maneuver, but risking being the ass who ends up blocking the crosswalk after miscalculating).

- First one, to stop too far away both from the road intersection and the pedestrian crossing. And of course you should not stop on top the pedestrian crossing. The car behind the Tesla, noticed that right away, and went over the Tesla. Even the Tesla driver in the video commented on that. I wonder what FSD version that driver has :-)

- Second failure, the car ahead, the one already over the road crossing started moving but the Tesla would not move at all. Only when the video author accelerated as he mentions in the video.

What light? There is none if you're talking about t=995.

Most intersections with unprotected left turns let you fit at least a whole car, sometimes two in the intersection to wait, and the third car has its front wheels on the paint, so it's totally in the intersection.

Even if the light for opposing traffic turns green while the turning car is still clearing the intersection, opposing traffic is legally required to wait and not enter the intersection until opposing traffic has cleared the intersection.

This extra liability would seem like making self driving cars error on the side of stopping to be in the interests of auto makers who would be exposed to that liability.

Imagine how crazy the law would be if this were the case: - Light changes from green to yellow 1ms before your front wheels pass the stop line - Other direction traffic runs a red light and hits you from the side - You're now somehow liable because your wheels entered the intersection with zero ability to react quickly enough (no human can react in 1ms and no car can stop that fast) and the other driver that clearly blew a red is not?

That would be pants on head crazy, tbh...

Not parituclarly relevant to the gridlock violation at issue here, but left on red from a one-way street to a one-way street is legal in the same conditions as right-on-red generally in a large majority of US states, a few of which also allow left-on-red from a two-way street to a one-way street. Left-on-red is only completely illegal in a few states.

Yeah, I've seen tickets written for this in Emeryville, CA for cars that entered on green and weren't going to clear the intersection in the next ten minutes. (I don't know if the intersection is still that bad.) Bicycle cop, because there was no way they'd get a car in there.

> Left on red is never legal

I recall that this was legal in Michigan, where I grew up, but it either had to be onto or off of a one way street. I could never remember which, so I didn't try it.

But I did see someone make a left on red on market street in SF from the rightmost lane, which is very illegal.

I believe it requires two one-way streets.

Although searching for this now, it does appear to also be allowed from a two-way,[0] which sounds vaguely familiar but isn't really the same as turning right on red.

[0] https://www.michigan.gov/msp/services/traffic-safety/traffic...

Regarding deceptive editing, plenty of people post their Teslas doing squirrely things and them intervening. So it's not like a secret that sometimes you have to intervene.

And, empirically, there's no evidence that these cars are less safe when driving this way. Tesla claims people driving with FSD enabled have 4x fewer accidents than those driving without it, and nobody has presented any data that disputes that.

"...Several attempts have been made to reach out to Tesla for comment over the years since these numbers first started coming out, however, Tesla closed its press relations office and no longer responds to press inquiries..."

Saying fsd on tela has the same statistic than the general driver population prints a grim picture, as it puts it in a strictly worse performance than peers vehicles (SUV or saloons depending on the model)

I am not joking when I say millions of times less information. The aggregate motor vehicle fatality rate in the US is ~1.3 per 100,000,000 miles. To even have a chance of proving safety you would need to analyze on the order of 1,000,000,000 miles exhaustively. In contrast, to demonstrate it is not safe you would only need a list of like 20 fatalitys over those same billion miles to affirmatively prove it is unsafe. A list of 20 names versus the raw data for 1,000,000,000 miles is at least a factor of millions in information required and qualitatively different to achieve. Proving safety is enormously difficult and basically requires direct access to the raw information. Proving unsafety is extremely easy and can be relatively easily demonstrated by third partys when dealing with safety critical systems.

Safety engineering is not some sort of new idea that we need to reinvent by the seat of our pants from first principles, it is a mature methodology regularly deployed in aerospace, medical, civil, and automotive engineering applications. Assumptions of safety are fundamentally incompatible with safety critical engineering. This is not a fringe opinion, it is literally a core concept at the foundations of safety critical engineering and not understanding that concept demonstrates a fundamental misunderstanding of modern safety. It is like a geologist arguing the Earth is flat; anybody who suggests such a idea is so fundamentally divorced from the ideas of the field that their knowledge not only fails to reach the level of a expert, they fail to reach the level of even a basic practitioner.

https://youtu.be/WowhH_Xry9s?t=272

See how is confusing the speed limit for cars and trucks, and doing a sudden break. More precisely, at 5:12.

So if you have car driving behind you, not expecting the sudden break, this could cause an accident. They are joking with people lives.

Ethics matter in Software Engineering.

"reckless" "untrained fanboys" "defective software" - what is this tone? Why is it reckless? Why do the fanboys need training? Why do you think the software is defective? These are significant unjustified claims!

To me, it seems each company has a different strategy for self-driving, which aren't directly comparable. Beta testing with relatively untrained people on public roads seems necessary for either strategy though.

California seems to think so.

>Beta testing with relatively untrained people on public roads seems necessary for either strategy though.

Then why is Tesla the only company doing it?

Mercedes has an L3 system shipping today and they didn't see any need to endanger my life to build it.

I can’t imagine letting it go for a full hour without causing some road rage or traffic blockage.

To be clear there is a definite driving “culture” in LA that is very permissive and aggressive (out of necessity). FSD doesn’t follow this culture.

And we can also test it in Ho Chi Minh. Tesla says it has superhuman abilities... - https://youtu.be/1ZupwFOhjl4

This is super annoying, but the most recent FSD update seems to have mostly fixed that.

It's still pretty crappy in a lot of other ways, but at least that improved.

I have read this comment before.

The problem is that issues may be fixed one week but re-appear later.

Self driving is not an exercise in fixing bugs despite what many people think.

So, they are getting what they need to make it better.

lmao do people do this for free?

You need to be able to accurately do bounding box detection in order to determine whether that billboard of a person is real or not or if that dog with a hat should be avoided.

Research has conclusively shown that vision only systems simply can't match LiDAR for this task.

And we also can move our eyes around in three dimensions to infer depth.

w.r.t. the super computer point, AI systems have been able to out perform humans for specialized tasks for a while.

This is another one, definitely one of my favorite examples ( at correct time) - https://youtu.be/2u6AgLuwVqI?t=86

Notice also on this example:

https://youtu.be/gzcR8RaC3-g?t=397

They have a driver attention detection system, that works even with dark glasses, as the driver in this case. Also even if you do not take over in the 10 seconds, like in the case for example of the driver fainting, or having a heart attack, the car will stop safely on the side of the road, alert emergency services, and unlock the doors for first responders.

Oh wait that is an unreasonably niche scenario to be worthy of consideration in the context of full autonomy.

https://www.carscoops.com/2022/05/tesla-driver-claims-model-...

Nobody else is doing that, and I resent their beta-testing features with the rest of us there as unwilling participants.